Table 1 Characteristics of the protein variant effect prediction tools assessed in this study. The table indicates their scoring ranges and thresholds, training data, summary information about features and, where applicable, machine learning method
Prediction tool | Score range | Deleterious score cutoff | Training data | Features | Machine learning method |
|---|---|---|---|---|---|
GERP++ | −12.0 to 6.17 | >0.047 | None | Infers conserved or constrained elements from 33 mammalian genomes | – |
fitCons | 0 to 1 | >0.4 | None | Functional genomics data mainly sourced from chromatin analysis, e.g. ChIP-seq, and evolutionary conservation data | – |
SIFT | 1 to 0 | <0.05 | None | Conservation data (MSA of homologous sequences) and transformed into normalised probability matrix | – |
PolyPhen | 0 to 1 | >0.5 | HumVar, HumDiv | Conservation data (MSA of homologous sequences), protein functional domain data and protein structural features | Naïve Bayes classifier |
CADD | 0 to 35+ | >15 | Simulated, Swissvar, HumVar | Integrates several annotations into a single score, e.g. SIFT, GERP++, PolyPhen, CPG distance, GC content | SVM |
Condel | 0 to 1 | >0.5 | Â | Builds a unified classification by integration output from a collection of tools, e.g. SIFT, PolyPhen | Weighted average normalised scores |
REVEL | 0 to 1 | >0.5 | HGMD, EPS | HGMD and rare EPS variants used for training | Random forest |
fathmm | 0 to 1 | >0.45 | HGMD, Swiss-Prot | Combines evolutionary conservation with disease-specific protein weights for intolerance to mutation | Hidden Markov models |